Reference
LLM API pricing, in one honest table
Provider list prices, as of 2026-07 — 30models, sorted cheapest to most expensive by input price. Rows we couldn’t verify against an official pricing page are marked ‡.
| Model | Org | Weights | Context | $ in / MTok | $ out / MTok | What it’s actually good at |
|---|---|---|---|---|---|---|
| Command R7B | Cohere | closed | 128K | $0.0375 | $0.1500 | ultra-cheap enterprise RAG and classification at scale |
| Llama 3.1 8B | Meta | open | 128K | $0.0500 | $0.1000 | cheap, fast, easily fine-tuned baseline for narrow tasks |
| Qwen3 32B | Alibaba | open | 131K (via YaRN) | $0.0800 | $0.2800 | strong reasoning-to-cost ratio in a single self-hostable dense model |
| Mistral Small 3‡ | Mistral AI | open | 128K | $0.0800 | $0.2000 | efficient general chat and instruction-following at a self-hostable size |
| Llama 4 Scout | Meta | open | 10M (advertised; effective recall shorter) | $0.1000 | $0.3000 | extremely long-context ingestion of codebases and long documents at low cost |
| GPT-4o mini | OpenAI | closed | 128K | $0.1500 | $0.6000 | cheap, fast, good-enough responses for high-volume simple tasks |
| Command R | Cohere | closed | 128K | $0.1500 | $0.6000 | efficient RAG and enterprise search-grounded generation |
| Llama 4 Maverick‡ | Meta | open | 1M | $0.2000 | $0.8500 | strong general chat and multimodal reasoning at open-weight cost |
| Qwen3 235B-A22B‡ | Alibaba | open | 131K (via YaRN) | $0.2000 | $0.6000 | flagship open-weight reasoning, coding, and agentic tool use with a toggleable thinking mode |
| Claude 3 Haiku | Anthropic | closed | 200K | $0.2500 | $1.25 | legacy cheap-tier chat and summarization |
| DeepSeek V3‡ | DeepSeek | open | 128K | $0.2700 | $1.10 | near-frontier general reasoning and coding at a fraction of closed-model cost |
| Gemini 2.5 Flash | closed | 1M | $0.3000 | $2.50 | high-throughput multimodal tasks needing large context at low cost | |
| Llama 3.3 70B | Meta | open | 128K | $0.3000 | $0.6000 | close-to-405B quality at 70B cost, a popular self-hosted workhorse |
| Qwen2.5 72B‡ | Alibaba | open | 128K | $0.3500 | $0.4000 | strong general and coding performance, still widely deployed |
| GLM-4.5‡ | Zhipu AI | open | 128K | $0.4000 | $1.60 | strong open-weight agentic coding and tool-use performance |
| DeepSeek R1‡ | DeepSeek | open | 128K | $0.5500 | $2.19 | step-by-step reasoning, math, and coding rivaling closed reasoning models at open-weight cost |
| Kimi K2‡ | Moonshot AI | open | 128K-262K | $0.5500 | $2.20 | frontier-competitive open-weight coding and agentic tool use at trillion-parameter scale |
| Claude Haiku 4.5‡ | Anthropic | closed | 200K | $1.00 | $5.00 | low-latency chat, classification, and simple extraction at scale |
| o4-mini‡ | OpenAI | closed | 200K | $1.10 | $4.40 | reasoning tasks at a fraction of o3's cost |
| Gemini 2.5 Pro‡ | closed | 1M | $1.25 | $10.00 | native multimodal reasoning and huge-context document or video analysis | |
| GPT-4.1‡ | OpenAI | closed | 1M | $2.00 | $8.00 | long-context comprehension and more reliable instruction following than GPT-4o |
| o3‡ | OpenAI | closed | 200K | $2.00 | $8.00 | hard multi-step math, science, and coding reasoning |
| Mistral Large 2‡ | Mistral AI | closed | 128K | $2.00 | $6.00 | strong multilingual reasoning and coding, a European-hosted alternative to US frontier models |
| GPT-4o‡ | OpenAI | closed | 128K | $2.50 | $10.00 | strong general-purpose chat, vision, and tool use with a good speed/quality balance |
| Command R+ | Cohere | closed | 128K | $2.50 | $10.00 | enterprise RAG, citations, and multi-step tool use with strong grounding |
| Llama 3.1 405B | Meta | open | 128K | $2.70 | $2.70 | frontier-adjacent open-weight reasoning and generation quality |
| Claude Sonnet 4.x‡ | Anthropic | closed | 200K | $3.00 | $15.00 | the default high-quality coding and reasoning workhorse most teams reach for first |
| Claude 3.5 Sonnet | Anthropic | closed | 200K | $3.00 | $15.00 | strong coding and agentic tool use, the model that popularized computer use |
| Grok 4‡ | xAI | closed | 256K | $3.00 | $15.00 | competitive frontier reasoning and coding, positioned against GPT-5 and Claude Opus |
| Claude Opus 4.x‡ | Anthropic | closed | 200K | $15.00 | $75.00 | deepest reasoning, careful long-form writing, and complex agentic coding |
‡ 18 of 30rows carry prices the source research couldn’t cross-confirm on an official provider pricing page as of 2026-07 — treat them as directional, not contractual. Price class (frontier, mid, budget, ultra-budget) is derived from list input price, not a vendor label. No benchmark scores are shown here because none of the sentences above are backed by a benchmark we ran ourselves — they’re read off the model card, not measured.
Why the sticker price isn’t the bill
What routing does to these numbers
Every price above is what you pay if every request hits that exact model. A router instead sends each request to the cheapest model that can still do the job, so most traffic never touches the frontier price at all. Here’s one worked example, computed from the table above — not a benchmark, just arithmetic.
500,000 requests/month, averaging 600 input and 400 output tokens each, asked for at Claude Sonnet 4.x list price, served instead by DeepSeek V3 — a right-sized model in a similar capability band.
- Claude Sonnet 4.x at list price
- $3,900/mo
- DeepSeek V3 serving the same volume
- $301.00/mo
- Delta
- −92%
This is one example at one volume, not a promise about your traffic — different prompts, task mixes, and quality bars change the answer. Run your own numbers on the cost calculator, or get an API key and see the routed price on your own requests. Full docs at /docs.